letta-nightly 0.11.6.dev20250902104140__py3-none-any.whl → 0.11.7.dev20250904045700__py3-none-any.whl

letta/schemas/tool_rule.py

@@ -20,6 +20,16 @@ class BaseToolRule(LettaBase):
         description="Optional Jinja2 template for generating agent prompt about this tool rule. Template can use variables like 'tool_name' and rule-specific attributes.",
     )
 
+    def __hash__(self):
+        """Base hash using tool_name and type."""
+        return hash((self.tool_name, self.type))
+
+    def __eq__(self, other):
+        """Base equality using tool_name and type."""
+        if not isinstance(other, BaseToolRule):
+            return False
+        return self.tool_name == other.tool_name and self.type == other.type
+
     def get_valid_tools(self, tool_call_history: List[str], available_tools: Set[str], last_function_response: Optional[str]) -> set[str]:
         raise NotImplementedError
 
@@ -54,6 +64,16 @@ class ChildToolRule(BaseToolRule):
         description="Optional Jinja2 template for generating agent prompt about this tool rule.",
     )
 
+    def __hash__(self):
+        """Hash including children list (sorted for consistency)."""
+        return hash((self.tool_name, self.type, tuple(sorted(self.children))))
+
+    def __eq__(self, other):
+        """Equality including children list."""
+        if not isinstance(other, ChildToolRule):
+            return False
+        return self.tool_name == other.tool_name and self.type == other.type and sorted(self.children) == sorted(other.children)
+
     def get_valid_tools(self, tool_call_history: List[str], available_tools: Set[str], last_function_response: Optional[str]) -> Set[str]:
         last_tool = tool_call_history[-1] if tool_call_history else None
         return set(self.children) if last_tool == self.tool_name else available_tools
@@ -71,6 +91,16 @@ class ParentToolRule(BaseToolRule):
         description="Optional Jinja2 template for generating agent prompt about this tool rule.",
     )
 
+    def __hash__(self):
+        """Hash including children list (sorted for consistency)."""
+        return hash((self.tool_name, self.type, tuple(sorted(self.children))))
+
+    def __eq__(self, other):
+        """Equality including children list."""
+        if not isinstance(other, ParentToolRule):
+            return False
+        return self.tool_name == other.tool_name and self.type == other.type and sorted(self.children) == sorted(other.children)
+
     def get_valid_tools(self, tool_call_history: List[str], available_tools: Set[str], last_function_response: Optional[str]) -> Set[str]:
         last_tool = tool_call_history[-1] if tool_call_history else None
         return set(self.children) if last_tool == self.tool_name else available_tools - set(self.children)
@@ -90,6 +120,24 @@ class ConditionalToolRule(BaseToolRule):
         description="Optional Jinja2 template for generating agent prompt about this tool rule.",
     )
 
+    def __hash__(self):
+        """Hash including all configuration fields."""
+        # convert dict to sorted tuple of items for consistent hashing
+        mapping_items = tuple(sorted(self.child_output_mapping.items()))
+        return hash((self.tool_name, self.type, self.default_child, mapping_items, self.require_output_mapping))
+
+    def __eq__(self, other):
+        """Equality including all configuration fields."""
+        if not isinstance(other, ConditionalToolRule):
+            return False
+        return (
+            self.tool_name == other.tool_name
+            and self.type == other.type
+            and self.default_child == other.default_child
+            and self.child_output_mapping == other.child_output_mapping
+            and self.require_output_mapping == other.require_output_mapping
+        )
+
     def get_valid_tools(self, tool_call_history: List[str], available_tools: Set[str], last_function_response: Optional[str]) -> Set[str]:
         """Determine valid tools based on function output mapping."""
         if not tool_call_history or tool_call_history[-1] != self.tool_name:
@@ -203,6 +251,16 @@ class MaxCountPerStepToolRule(BaseToolRule):
         description="Optional Jinja2 template for generating agent prompt about this tool rule.",
     )
 
+    def __hash__(self):
+        """Hash including max_count_limit."""
+        return hash((self.tool_name, self.type, self.max_count_limit))
+
+    def __eq__(self, other):
+        """Equality including max_count_limit."""
+        if not isinstance(other, MaxCountPerStepToolRule):
+            return False
+        return self.tool_name == other.tool_name and self.type == other.type and self.max_count_limit == other.max_count_limit
+
     def get_valid_tools(self, tool_call_history: List[str], available_tools: Set[str], last_function_response: Optional[str]) -> Set[str]:
         """Restricts the tool if it has been called max_count_limit times in the current step."""
         count = tool_call_history.count(self.tool_name)
@@ -214,6 +272,18 @@ class MaxCountPerStepToolRule(BaseToolRule):
         return available_tools
 
 
+class RequiresApprovalToolRule(BaseToolRule):
+    """
+    Represents a tool rule configuration which requires approval before the tool can be invoked.
+    """
+
+    type: Literal[ToolRuleType.requires_approval] = ToolRuleType.requires_approval
+
+    def get_valid_tools(self, tool_call_history: List[str], available_tools: Set[str], last_function_response: Optional[str]) -> Set[str]:
+        """Does not enforce any restrictions on which tools are valid"""
+        return available_tools
+
+
 ToolRule = Annotated[
     Union[
         ChildToolRule,
@@ -224,6 +294,7 @@ ToolRule = Annotated[
         RequiredBeforeExitToolRule,
         MaxCountPerStepToolRule,
         ParentToolRule,
+        RequiresApprovalToolRule,
     ],
     Field(discriminator="type"),
 ]

letta/serialize_schemas/marshmallow_agent.py

@@ -5,8 +5,7 @@ from sqlalchemy import func
 from sqlalchemy.orm import sessionmaker
 
 import letta
-from letta.orm import Agent
-from letta.orm import Message as MessageModel
+from letta.orm import Agent, Message as MessageModel
 from letta.schemas.agent import AgentState as PydanticAgentState
 from letta.schemas.user import User
 from letta.serialize_schemas.marshmallow_agent_environment_variable import SerializedAgentEnvironmentVariableSchema

letta/server/rest_api/app.py

@@ -261,7 +261,7 @@ def create_application() -> "FastAPI":
 
     @app.exception_handler(BedrockPermissionError)
     async def bedrock_permission_error_handler(request, exc: BedrockPermissionError):
-        logger.error(f"Bedrock permission denied.")
+        logger.error("Bedrock permission denied.")
         if SENTRY_ENABLED:
             sentry_sdk.capture_exception(exc)
 
@@ -433,10 +433,10 @@ def start_server(
         if IS_WINDOWS:
             # Windows doesn't those the fancy unicode characters
             print(f"Server running at: http://{host or 'localhost'}:{port or REST_DEFAULT_PORT}")
-            print(f"View using ADE at: https://app.letta.com/development-servers/local/dashboard\n")
+            print("View using ADE at: https://app.letta.com/development-servers/local/dashboard\n")
         else:
             print(f"▶ Server running at: http://{host or 'localhost'}:{port or REST_DEFAULT_PORT}")
-            print(f"▶ View using ADE at: https://app.letta.com/development-servers/local/dashboard\n")
+            print("▶ View using ADE at: https://app.letta.com/development-servers/local/dashboard\n")
 
         if importlib.util.find_spec("granian") is not None and settings.use_granian:
             # Experimental Granian engine

letta/server/rest_api/auth/index.py

@@ -22,7 +22,6 @@ class AuthRequest(BaseModel):
 
 
 def setup_auth_router(server: SyncServer, interface: QueuingInterface, password: str) -> APIRouter:
-
     @router.post("/auth", tags=["auth"], response_model=AuthResponse)
     def authenticate_user(request: AuthRequest) -> AuthResponse:
         """

letta/server/rest_api/interface.py

@@ -377,9 +377,9 @@ class StreamingServerInterface(AgentChunkStreamingInterface):
     ):
         """Add an item to the deque"""
         assert self._active, "Generator is inactive"
-        assert (
-            isinstance(item, LettaMessage) or isinstance(item, LegacyLettaMessage) or isinstance(item, MessageStreamStatus)
-        ), f"Wrong type: {type(item)}"
+        assert isinstance(item, LettaMessage) or isinstance(item, LegacyLettaMessage) or isinstance(item, MessageStreamStatus), (
+            f"Wrong type: {type(item)}"
+        )
 
         self._chunks.append(item)
         self._event.set()  # Signal that new data is available
@@ -731,13 +731,11 @@ class StreamingServerInterface(AgentChunkStreamingInterface):
 
                     # If we have main_json, we should output a ToolCallMessage
                     elif updates_main_json:
-
                         # If there's something in the function_name buffer, we should release it first
                         # NOTE: we could output it as part of a chunk that has both name and args,
                         #       however the frontend may expect name first, then args, so to be
                         #       safe we'll output name first in a separate chunk
                         if self.function_name_buffer:
-
                             # use_assisitant_message means that we should also not release main_json raw, and instead should only release the contents of "message": "..."
                             if self.use_assistant_message and self.function_name_buffer == self.assistant_message_tool_name:
                                 processed_chunk = None
@@ -778,7 +776,6 @@ class StreamingServerInterface(AgentChunkStreamingInterface):
                         # If there was nothing in the name buffer, we can proceed to
                         # output the arguments chunk as a ToolCallMessage
                         else:
-
                             # use_assisitant_message means that we should also not release main_json raw, and instead should only release the contents of "message": "..."
                             if self.use_assistant_message and (
                                 self.last_flushed_function_name is not None
@@ -860,7 +857,6 @@ class StreamingServerInterface(AgentChunkStreamingInterface):
                                     # clear buffers
                                     self.function_id_buffer = None
                             else:
-
                                 # There may be a buffer from a previous chunk, for example
                                 # if the previous chunk had arguments but we needed to flush name
                                 if self.function_args_buffer:
@@ -997,7 +993,6 @@ class StreamingServerInterface(AgentChunkStreamingInterface):
             # Otherwise, do simple chunks of ToolCallMessage
 
             else:
-
                 tool_call_delta = {}
                 if tool_call.id:
                     tool_call_delta["id"] = tool_call.id
@@ -1073,7 +1068,6 @@ class StreamingServerInterface(AgentChunkStreamingInterface):
             tool_call = message_delta.tool_calls[0]
 
             if tool_call.function:
-
                 # Track the function name while streaming
                 # If we were previously on a 'send_message', we need to 'toggle' into 'content' mode
                 if tool_call.function.name:
@@ -1154,7 +1148,6 @@ class StreamingServerInterface(AgentChunkStreamingInterface):
     def internal_monologue(self, msg: str, msg_obj: Optional[Message] = None, chunk_index: Optional[int] = None):
         """Letta generates some internal monologue"""
         if not self.streaming_mode:
-
             # create a fake "chunk" of a stream
             # processed_chunk = {
             #     "internal_monologue": msg,
@@ -1268,7 +1261,6 @@ class StreamingServerInterface(AgentChunkStreamingInterface):
                             print(f"Failed to parse function message: {e}")
 
                 else:
-
                     try:
                         func_args = parse_json(function_call.function.arguments)
                     except:

letta/server/rest_api/redis_stream_manager.py

@@ -140,9 +140,7 @@ class RedisSSEStreamWriter:
 
             self.last_flush[run_id] = time.time()
 
-            logger.debug(
-                f"Flushed {len(chunks)} chunks to Redis stream {stream_key}, " f"seq_ids {chunks[0]['seq_id']}-{chunks[-1]['seq_id']}"
-            )
+            logger.debug(f"Flushed {len(chunks)} chunks to Redis stream {stream_key}, seq_ids {chunks[0]['seq_id']}-{chunks[-1]['seq_id']}")
 
             if chunks[-1].get("complete") == "true":
                 self._cleanup_run(run_id)
@@ -227,7 +225,8 @@ async def create_background_stream_processor(
     except Exception as e:
         logger.error(f"Error processing stream for run {run_id}: {e}")
         # Write error chunk
-        error_chunk = {"error": {"message": str(e)}}
+        # error_chunk = {"error": {"message": str(e)}}
+        error_chunk = {"error": str(e), "code": "INTERNAL_SERVER_ERROR"}
         await writer.write_chunk(run_id=run_id, data=f"event: error\ndata: {json.dumps(error_chunk)}\n\n", is_complete=True)
     finally:
         if should_stop_writer:

letta/server/rest_api/routers/v1/agents.py

@@ -2,7 +2,7 @@ import asyncio
 import json
 import traceback
 from datetime import datetime, timezone
-from typing import Annotated, Any, Dict, List, Optional, Union
+from typing import Annotated, Any, Dict, List, Literal, Optional, Union
 
 from fastapi import APIRouter, Body, Depends, File, Form, Header, HTTPException, Query, Request, UploadFile, status
 from fastapi.responses import JSONResponse
@@ -32,9 +32,15 @@ from letta.schemas.job import JobStatus, JobUpdate, LettaRequestConfig
 from letta.schemas.letta_message import LettaMessageUnion, LettaMessageUpdateUnion, MessageType
 from letta.schemas.letta_request import LettaAsyncRequest, LettaRequest, LettaStreamingRequest
 from letta.schemas.letta_response import LettaResponse
-from letta.schemas.memory import ContextWindowOverview, CreateArchivalMemory, Memory
+from letta.schemas.memory import (
+    ArchivalMemorySearchResponse,
+    ArchivalMemorySearchResult,
+    ContextWindowOverview,
+    CreateArchivalMemory,
+    Memory,
+)
 from letta.schemas.message import MessageCreate
-from letta.schemas.passage import Passage, PassageUpdate
+from letta.schemas.passage import Passage
 from letta.schemas.run import Run
 from letta.schemas.source import Source
 from letta.schemas.tool import Tool
@@ -155,8 +161,8 @@ async def export_agent_serialized(
     server: "SyncServer" = Depends(get_letta_server),
     actor_id: str | None = Header(None, alias="user_id"),
     use_legacy_format: bool = Query(
-        True,
-        description="If true, exports using the legacy single-agent format. If false, exports using the new multi-entity format.",
+        False,
+        description="If true, exports using the legacy single-agent format (v1). If false, exports using the new multi-entity format (v2).",
     ),
     # do not remove, used to autogeneration of spec
     # TODO: Think of a better way to export AgentFileSchema
@@ -252,6 +258,7 @@ async def import_agent(
     project_id: str | None = None,
     strip_messages: bool = False,
     env_vars: Optional[dict[str, Any]] = None,
+    override_embedding_handle: Optional[str] = None,
 ) -> List[str]:
     """
     Import an agent using the new AgentFileSchema format.
@@ -262,12 +269,19 @@ async def import_agent(
         raise HTTPException(status_code=422, detail=f"Invalid agent file schema: {e!s}")
 
     try:
+        if override_embedding_handle:
+            embedding_config_override = await server.get_cached_embedding_config_async(actor=actor, handle=override_embedding_handle)
+        else:
+            embedding_config_override = None
+
         import_result = await server.agent_serialization_manager.import_file(
             schema=agent_schema,
             actor=actor,
             append_copy_suffix=append_copy_suffix,
             override_existing_tools=override_existing_tools,
             env_vars=env_vars,
+            override_embedding_config=embedding_config_override,
+            project_id=project_id,
         )
 
         if not import_result.success:
@@ -296,11 +310,16 @@ async def import_agent_serialized(
     file: UploadFile = File(...),
     server: "SyncServer" = Depends(get_letta_server),
     actor_id: str | None = Header(None, alias="user_id"),
+    x_override_embedding_model: str | None = Header(None, alias="x-override-embedding-model"),
     append_copy_suffix: bool = Form(True, description='If set to True, appends "_copy" to the end of the agent name.'),
     override_existing_tools: bool = Form(
         True,
         description="If set to True, existing tools can get their source code overwritten by the uploaded tool definitions. Note that Letta core tools can never be updated externally.",
     ),
+    override_embedding_handle: Optional[str] = Form(
+        None,
+        description="Override import with specific embedding handle.",
+    ),
     project_id: str | None = Form(None, description="The project ID to associate the uploaded agent with."),
     strip_messages: bool = Form(
         False,
@@ -333,6 +352,9 @@ async def import_agent_serialized(
         if not isinstance(env_vars, dict):
             raise HTTPException(status_code=400, detail="env_vars_json must be a valid JSON string")
 
+    # Prioritize header over form data for override_embedding_handle
+    final_override_embedding_handle = x_override_embedding_model or override_embedding_handle
+
     # Check if the JSON is AgentFileSchema or AgentSchema
     # TODO: This is kind of hacky, but should work as long as dont' change the schema
     if "agents" in agent_json and isinstance(agent_json.get("agents"), list):
@@ -346,6 +368,7 @@ async def import_agent_serialized(
             project_id=project_id,
             strip_messages=strip_messages,
             env_vars=env_vars,
+            override_embedding_handle=final_override_embedding_handle,
         )
     else:
         # This is a legacy AgentSchema
@@ -464,6 +487,25 @@ async def detach_tool(
     return await server.agent_manager.get_agent_by_id_async(agent_id=agent_id, actor=actor)
 
 
+@router.patch("/{agent_id}/tools/approval/{tool_name}", response_model=AgentState, operation_id="modify_approval")
+async def modify_approval(
+    agent_id: str,
+    tool_name: str,
+    requires_approval: bool,
+    server: "SyncServer" = Depends(get_letta_server),
+    actor_id: str | None = Header(None, alias="user_id"),
+):
+    """
+    Attach a tool to an agent.
+    """
+    actor = await server.user_manager.get_actor_or_default_async(actor_id=actor_id)
+    await server.agent_manager.modify_approvals_async(
+        agent_id=agent_id, tool_name=tool_name, requires_approval=requires_approval, actor=actor
+    )
+    # TODO: Unfortunately we need this to preserve our current API behavior
+    return await server.agent_manager.get_agent_by_id_async(agent_id=agent_id, actor=actor)
+
+
 @router.patch("/{agent_id}/sources/attach/{source_id}", response_model=AgentState, operation_id="attach_source_to_agent")
 async def attach_source(
     agent_id: str,
@@ -937,22 +979,62 @@ async def create_passage(
     """
     actor = await server.user_manager.get_actor_or_default_async(actor_id=actor_id)
 
-    return await server.insert_archival_memory_async(agent_id=agent_id, memory_contents=request.text, actor=actor)
+    return await server.insert_archival_memory_async(
+        agent_id=agent_id, memory_contents=request.text, actor=actor, tags=request.tags, created_at=request.created_at
+    )
 
 
-@router.patch("/{agent_id}/archival-memory/{memory_id}", response_model=list[Passage], operation_id="modify_passage")
-def modify_passage(
+@router.get("/{agent_id}/archival-memory/search", response_model=ArchivalMemorySearchResponse, operation_id="search_archival_memory")
+async def search_archival_memory(
     agent_id: str,
-    memory_id: str,
-    passage: PassageUpdate = Body(...),
+    query: str = Query(..., description="String to search for using semantic similarity"),
+    tags: Optional[List[str]] = Query(None, description="Optional list of tags to filter search results"),
+    tag_match_mode: Literal["any", "all"] = Query(
+        "any", description="How to match tags - 'any' to match passages with any of the tags, 'all' to match only passages with all tags"
+    ),
+    top_k: Optional[int] = Query(None, description="Maximum number of results to return. Uses system default if not specified"),
+    start_datetime: Optional[datetime] = Query(None, description="Filter results to passages created after this datetime"),
+    end_datetime: Optional[datetime] = Query(None, description="Filter results to passages created before this datetime"),
     server: "SyncServer" = Depends(get_letta_server),
-    actor_id: str | None = Header(None, alias="user_id"),  # Extract user_id from header, default to None if not present
+    actor_id: str | None = Header(None, alias="user_id"),
 ):
     """
-    Modify a memory in the agent's archival memory store.
+    Search archival memory using semantic (embedding-based) search with optional temporal filtering.
+
+    This endpoint allows manual triggering of archival memory searches, enabling users to query
+    an agent's archival memory store directly via the API. The search uses the same functionality
+    as the agent's archival_memory_search tool but is accessible for external API usage.
     """
-    actor = server.user_manager.get_user_or_default(user_id=actor_id)
-    return server.modify_archival_memory(agent_id=agent_id, memory_id=memory_id, passage=passage, actor=actor)
+    actor = await server.user_manager.get_actor_or_default_async(actor_id=actor_id)
+
+    try:
+        # convert datetime to string in ISO 8601 format
+        start_datetime = start_datetime.isoformat() if start_datetime else None
+        end_datetime = end_datetime.isoformat() if end_datetime else None
+
+        # Use the shared agent manager method
+        formatted_results, count = await server.agent_manager.search_agent_archival_memory_async(
+            agent_id=agent_id,
+            actor=actor,
+            query=query,
+            tags=tags,
+            tag_match_mode=tag_match_mode,
+            top_k=top_k,
+            start_datetime=start_datetime,
+            end_datetime=end_datetime,
+        )
+
+        # Convert to proper response schema
+        search_results = [ArchivalMemorySearchResult(**result) for result in formatted_results]
+
+        return ArchivalMemorySearchResponse(results=search_results, count=count)
+
+    except NoResultFound as e:
+        raise HTTPException(status_code=404, detail=f"Agent with id={agent_id} not found for user_id={actor.id}.")
+    except ValueError as e:
+        raise HTTPException(status_code=400, detail=str(e))
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=f"Internal server error during archival memory search: {str(e)}")
 
 
 # TODO(ethan): query or path parameter for memory_id?
@@ -1049,6 +1131,8 @@ async def send_message(
     Process a user message and return the agent's response.
     This endpoint accepts a message from a user and processes it through the agent.
     """
+    if len(request.messages) == 0:
+        raise ValueError("Messages must not be empty")
     request_start_timestamp_ns = get_utc_timestamp_ns()
     MetricRegistry().user_message_counter.add(1, get_ctx_attributes())
 
@@ -1067,6 +1151,7 @@ async def send_message(
         "azure",
         "xai",
         "groq",
+        "deepseek",
     ]
 
     # Create a new run for execution tracking
@@ -1197,6 +1282,9 @@ async def send_message_streaming(
     request_start_timestamp_ns = get_utc_timestamp_ns()
     MetricRegistry().user_message_counter.add(1, get_ctx_attributes())
 
+    # TODO (cliandy): clean this up
+    redis_client = await get_redis_client()
+
     actor = await server.user_manager.get_actor_or_default_async(actor_id=actor_id)
     # TODO: This is redundant, remove soon
     agent = await server.agent_manager.get_agent_by_id_async(agent_id, actor, include_relationships=["multi_agent_group"])
@@ -1212,8 +1300,9 @@ async def send_message_streaming(
         "azure",
         "xai",
         "groq",
+        "deepseek",
     ]
-    model_compatible_token_streaming = agent.llm_config.model_endpoint_type in ["anthropic", "openai", "bedrock"]
+    model_compatible_token_streaming = agent.llm_config.model_endpoint_type in ["anthropic", "openai", "bedrock", "deepseek"]
 
     # Create a new job for execution tracking
     if settings.track_agent_run:
@@ -1236,14 +1325,11 @@ async def send_message_streaming(
             ),
             actor=actor,
         )
+        job_update_metadata = None
+        await redis_client.set(f"{REDIS_RUN_ID_PREFIX}:{agent_id}", run.id if run else None)
     else:
         run = None
 
-    job_update_metadata = None
-    # TODO (cliandy): clean this up
-    redis_client = await get_redis_client()
-    await redis_client.set(f"{REDIS_RUN_ID_PREFIX}:{agent_id}", run.id if run else None)
-
     try:
         if agent_eligible and model_compatible:
             if agent.enable_sleeptime and agent.agent_type != AgentType.voice_convo_agent:
@@ -1281,6 +1367,23 @@ async def send_message_streaming(
                     ),
                 )
 
+            if request.stream_tokens and model_compatible_token_streaming:
+                raw_stream = agent_loop.step_stream(
+                    input_messages=request.messages,
+                    max_steps=request.max_steps,
+                    use_assistant_message=request.use_assistant_message,
+                    request_start_timestamp_ns=request_start_timestamp_ns,
+                    include_return_message_types=request.include_return_message_types,
+                )
+            else:
+                raw_stream = agent_loop.step_stream_no_tokens(
+                    request.messages,
+                    max_steps=request.max_steps,
+                    use_assistant_message=request.use_assistant_message,
+                    request_start_timestamp_ns=request_start_timestamp_ns,
+                    include_return_message_types=request.include_return_message_types,
+                )
+
             from letta.server.rest_api.streaming_response import StreamingResponseWithStatusCode, add_keepalive_to_stream
 
             if request.background and settings.track_agent_run:
@@ -1294,23 +1397,6 @@ async def send_message_streaming(
                         ),
                     )
 
-                if request.stream_tokens and model_compatible_token_streaming:
-                    raw_stream = agent_loop.step_stream(
-                        input_messages=request.messages,
-                        max_steps=request.max_steps,
-                        use_assistant_message=request.use_assistant_message,
-                        request_start_timestamp_ns=request_start_timestamp_ns,
-                        include_return_message_types=request.include_return_message_types,
-                    )
-                else:
-                    raw_stream = agent_loop.step_stream_no_tokens(
-                        request.messages,
-                        max_steps=request.max_steps,
-                        use_assistant_message=request.use_assistant_message,
-                        request_start_timestamp_ns=request_start_timestamp_ns,
-                        include_return_message_types=request.include_return_message_types,
-                    )
-
                 asyncio.create_task(
                     create_background_stream_processor(
                         stream_generator=raw_stream,
@@ -1319,55 +1405,21 @@ async def send_message_streaming(
                     )
                 )
 
-                stream = redis_sse_stream_generator(
+                raw_stream = redis_sse_stream_generator(
                     redis_client=redis_client,
                     run_id=run.id,
                 )
 
-                if request.include_pings and settings.enable_keepalive:
-                    stream = add_keepalive_to_stream(stream, keepalive_interval=settings.keepalive_interval)
-
-                return StreamingResponseWithStatusCode(
-                    stream,
-                    media_type="text/event-stream",
-                )
-
-            if request.stream_tokens and model_compatible_token_streaming:
-                raw_stream = agent_loop.step_stream(
-                    input_messages=request.messages,
-                    max_steps=request.max_steps,
-                    use_assistant_message=request.use_assistant_message,
-                    request_start_timestamp_ns=request_start_timestamp_ns,
-                    include_return_message_types=request.include_return_message_types,
-                )
-                # Conditionally wrap with keepalive based on request parameter
-                if request.include_pings and settings.enable_keepalive:
-                    stream = add_keepalive_to_stream(raw_stream, keepalive_interval=settings.keepalive_interval)
-                else:
-                    stream = raw_stream
-
-                result = StreamingResponseWithStatusCode(
-                    stream,
-                    media_type="text/event-stream",
-                )
+            # Conditionally wrap with keepalive based on request parameter
+            if request.include_pings and settings.enable_keepalive:
+                stream = add_keepalive_to_stream(raw_stream, keepalive_interval=settings.keepalive_interval)
             else:
-                raw_stream = agent_loop.step_stream_no_tokens(
-                    request.messages,
-                    max_steps=request.max_steps,
-                    use_assistant_message=request.use_assistant_message,
-                    request_start_timestamp_ns=request_start_timestamp_ns,
-                    include_return_message_types=request.include_return_message_types,
-                )
-                # Conditionally wrap with keepalive based on request parameter
-                if request.include_pings and settings.enable_keepalive:
-                    stream = add_keepalive_to_stream(raw_stream, keepalive_interval=settings.keepalive_interval)
-                else:
-                    stream = raw_stream
-
-                result = StreamingResponseWithStatusCode(
-                    stream,
-                    media_type="text/event-stream",
-                )
+                stream = raw_stream
+
+            result = StreamingResponseWithStatusCode(
+                stream,
+                media_type="text/event-stream",
+            )
         else:
             result = await server.send_message_to_agent(
                 agent_id=agent_id,
@@ -1382,11 +1434,13 @@ async def send_message_streaming(
                 request_start_timestamp_ns=request_start_timestamp_ns,
                 include_return_message_types=request.include_return_message_types,
             )
-        job_status = JobStatus.running
+        if settings.track_agent_run:
+            job_status = JobStatus.running
         return result
     except Exception as e:
-        job_update_metadata = {"error": str(e)}
-        job_status = JobStatus.failed
+        if settings.track_agent_run:
+            job_update_metadata = {"error": str(e)}
+            job_status = JobStatus.failed
         raise
     finally:
         if settings.track_agent_run:
@@ -1469,7 +1523,10 @@ async def _process_message_background(
             "google_vertex",
             "bedrock",
             "ollama",
+            "azure",
+            "xai",
             "groq",
+            "deepseek",
         ]
         if agent_eligible and model_compatible:
             if agent.enable_sleeptime and agent.agent_type != AgentType.voice_convo_agent:
@@ -1660,6 +1717,7 @@ async def preview_raw_payload(
         "azure",
         "xai",
         "groq",
+        "deepseek",
     ]
 
     if agent_eligible and model_compatible:
@@ -1731,6 +1789,7 @@ async def summarize_agent_conversation(
         "azure",
         "xai",
         "groq",
+        "deepseek",
     ]
 
     if agent_eligible and model_compatible:

letta/server/rest_api/routers/v1/blocks.py

@@ -34,7 +34,7 @@ async def list_blocks(
     ),
     label_search: Optional[str] = Query(
         None,
-        description=("Search blocks by label. If provided, returns blocks that match this label. " "This is a full-text search on labels."),
+        description=("Search blocks by label. If provided, returns blocks that match this label. This is a full-text search on labels."),
     ),
     description_search: Optional[str] = Query(
         None,